Research On SAR Image Ship Detection Algorithm Based On Deep Learning

At present,deep learning algorithms have been widely used in ship detection in SAR(Synthetic Aperture Radar)images.Ships in SAR images are sparsely distributed,have large size differences and large aspect ratios.Therefore,the Faster R-CNN algo-rithm has problems such as low accuracy and high missed detection rate when detecting this type of ships.To address this issue,the main work carried out in this thesis is as follows:First of all,aiming at the limitations of the ship detection algorithms in SAR images based on deep learning in practical applications,a new ship detection network structure is proposed.The Guided Anchoring-RPN structure is adopted in the new ship target de-tection network structure to increase the detection accuracy for ships with very different aspect ratios.And Combining the advantages of the convolutional layer and the fully con-nected layer in positioning and classification,the new ship detection network performs the tasks of ship position regression and classification respectively.These structures reduce the false detection rate of ships,and can also improve the wrapping of the predicted boxes frame to the ships.Aiming at the problem of small-sized ships missed alarms,the multi-scale fusion feature map generated by the feature pyramid network is used for classifica-tion and regression.And RoIAlign can reduce the quantization error generated during the training process,thereby reducing the probability of missed inspections of small ships.In terms of generalization performance,the proposed detection network has the highest ac-curacy on the SSDD data set,and its generalization performance is better.Secondly,the new detection network,Faster R-CNN detection network and SSD de-tection network are three detection networks that are trained on the open source data set SAR-Ship-Dataset.The trained networks perform comparative experiments and analysis from the detection accuracy,the pr curve of anchors,and the robustness.Compared with Faster R-CNN detection network and SSD detection network,the new detection network has a performance improvement of 6.6% and 10.9%.Judging from the PR curve of pre-dicted bounding boxes,the proposed detection network has higher detection accuracy and lower missed alarm rate.In terms of robustness,the detection performance of the pro-posed detection network under four different types of noise is better than the other two detection networks.In summary,for ships with small size and wide disparity in aspect ratio,the proposed detection network has better detection performance,lower probability of missed detection,stronger robustness and better generalization performance than the other two networks.